harshad317/ICSR_dataset

ICSR Detection Dataset (binary, balanced, with entities)

A balanced English dataset for binary Individual Case Safety Report (ICSR) detection in biomedical literature. Each row is a PubMed / Europe PMC title + abstract, labelled for whether it constitutes an individual case safety report (a suspected drug/poison → adverse reaction in an identifiable patient), together with the entities that drove the decision.

Intended as a first-pass triage / screening corpus for pharmacovigilance literature.

Fields

column	type	description
text	string	"Title\n\nAbstract" (canonicalised; see Text normalisation)
entities	string (JSON)	{"patient": [...], "suspect_drug": [...], "adverse_reaction": [...]} — the extracted spans. Empty lists for most DISCARDs (a DISCARD failed a criterion, so there is nothing to extract).
label	string	"ICSR" or "DISCARD"

import json
from datasets import load_dataset

ds = load_dataset("harshad317/ICSR_dataset")
row = ds["train"][0]
print(row["label"], row["text"][:80])
print(json.loads(row["entities"]))   # entities is a JSON string

Splits

Balanced exactly 50/50 in every split (majority class downsampled, then stratified per class, seed 42). Split ratio 80 / 10 / 10.

split	rows	ICSR	DISCARD
train	128,370	64,185	64,185
validation	16,046	8,023	8,023
test	16,048	8,024	8,024
total	160,464	80,232	80,232

Labeling definition

An article is labelled ICSR only if all four hold for the case the article reports firsthand:

1. Identifiable patient — an identifiable individual (age / sex / initials / individually described case). Case series of individually described patients count. Aggregate cohorts ("n=120 patients", randomized trials reporting pooled results) do not.
2. Suspect agent — a suspected drug or poison administered to / taken by that patient and presented as possibly involved. Medical devices are out of scope; poisons / toxic substances (pesticide, venom, injected oils) are in scope. A pathogen by itself (an infection) is not a suspect agent.
3. Adverse event — a harmful/unintended occurrence in that patient. The underlying disease, disease progression, and lack of efficacy do not count.
4. Causal association — the event is attributed to the suspect agent (suspected agent → reaction link), not mere co-occurrence. A drug used to treat the event is not a suspect drug.

Otherwise the article is DISCARD.

How it was built

The positive and negative classes were pooled from LLM labelling pipelines run over PubMed / Europe PMC literature, deduplicated by normalised title, and balanced 50/50.

Two labelling pipelines contributed labels:

• Entity-extraction + deterministic rule (GPT-5.4, high reasoning): extract patient / suspect drug / adverse event / causal link, then apply the four-criteria rule above.
• Decomposed relabelling: cheap per-entity extraction (GPT-5.4-nano) → strong causal judge (GPT-5.4, high reasoning) → deterministic rule, with confidence-based routing. Rows below a confidence threshold were quarantined and excluded, not guessed.

Negatives include a deliberately large share of hard negatives — drug-related biomedical literature that is not an ICSR — to force content discrimination rather than keyword matching.

Text normalisation

Source pipelines stored text in different formats ("title: X\n\nastract: Y", "Title: X\n\nAbtract: Y", and raw "Title\n\nAbstract"). All rows were normalised to a single canonical "Title\n\nAbstract" form. Without this the prefix format correlated with the label source, letting a classifier shortcut on formatting instead of content.

⚠️ Limitations — read before using

• The labels are LLM-generated and NOT validated against a human gold standard. Their accuracy is unmeasured. Treat them as machine-generated annotations, not ground truth.
• Two labelling pipelines contributed labels and disagree on a non-trivial fraction of articles. ~4,900 articles were labelled ICSR by one pipeline and DISCARD by the other; these conflicts were resolved ICSR-wins (recall-leaning), so some are likely false positives.
• Same-family dependence. All labels derive from OpenAI models; shared systematic blind spots will not have been caught.
• Known error modes (not fully mitigated): the small model's suspect vs treatment role call (the largest driver of label changes); recall misses on case series; pathogens mistyped as suspect drugs.
• Uncertain rows were dropped, not resolved — the retained set is easier than the real-world distribution, so models may underperform on genuinely ambiguous cases (the ones that matter most in pharmacovigilance).
• Balance is artificial. The natural ICSR rate is far below 50%; precision on this balanced test split will overstate deployment precision on an unbalanced stream.
• Deduplication is title-based. Near-duplicate articles with differing titles may appear across splits and could inflate test scores.
• Truncation. Very long documents were truncated during labelling.
• English biomedical titles + abstracts only.

Intended use

• First-pass filter / triage for pharmacovigilance literature screening.
• Biomedical NLP research on adverse-event, causality, and entity extraction.

Out of scope

• Not a medical device. Not a regulatory or clinical determination. Do not use as the sole basis for safety reporting or clinical decisions; keep a qualified human in the loop.
• Given the unvalidated labels, do not treat this as a benchmark of record without first establishing a human-annotated reference set.

Source data

Derived from openly available PubMed / Europe PMC titles and abstracts (published biomedical literature; no patient-identifying data beyond what appears in published case reports).